The Future of Code Generation | Guillermo Rauch, CEO of Vercel | Ep. 20

You've been coding for hours and hours and he didn't even know what the output was. He was just impressed by the fact that someone could be so locked in. And so I think programming taught me that. It taught me how to focus. It taught me to be disciplined. It taught me to receive this negative feedback from the compiler and overcome it. I do think, you know, we'll need to find what the next version of that is because I don't think it's going to be programming necessarily.

All right, Guillermo, thank you so much for making the time for this. I'm really excited to chat with you today.

People are excited on X as well, so.

So I want to get into CodeGen, uh, but before we go there, can you talk about what you worked on before Vercel, you know, with Next.js, other projects, and maybe how that fed into what you've built at Vercel?

Yeah, I had a startup before Vercel that I exited to Automattic, the parent company of WordPress.com. And it was quite a successful journey for me because it was my first startup. It's nice to have an exit, but also one of the meta things that I learned, I was a CTO. One of the meta things that I learned is as a CTO, how can you influence your team in the best possible way? Like they're going to engineer the right things, they're going to have the best tools. The one thing I did that was revolutionary for my team was spending a lot of time in getting the CI/CD process, meaning continuous integration and continuous deployment of the code that they would write, get it as efficient as possible. Meaning you write a little feature, you push code to Git, You get a URL back. I built a real-time system. My background was in writing real-time frameworks, so I was obsessed with real-time streaming of data. And so people would push to Git and then I would give them this URL that had a commit ID, dot, my company was called LearnBoost. I think it was like learnboostdemo.com. And so imagine that you're almost editing the internet in real time. That was the feeling that I wanted to give my employees. And obviously I did a lot of other things. I chose technology stacks and whatever, but when I would ask my colleagues like, what was the thing that was most impactful in your time here? It was that iteration velocity, the deployment velocity, the tooling being really neatly configured. I always do the exercise in my head of you show up to a new company. Isn't it nice that Apple kind of figured out operating system and hardware? And I give you a new laptop, actually, you probably— I don't know about you, but I look forward to having a new laptop. It's the best. It's so new and nice. Unbelievable. And that's only because they figured out this bootstrapping problem of like everything is ready to be used. I wanted to give that feeling, but for your development tools to my team.

Versus you get a new, you know, you're setting up a whole environment and it takes forever.

Yeah, install these 20 million things, fight the tools, and then maybe you get to work on something. Yeah. And so that was by far the most impactful thing that I did. And I wanted to, you know, it was sort of subconscious at the time, but essentially I wanted to turn that into a company. I couldn't wait to, you know, you have that insight because then I kind of saw that not every company had that. WordPress was really good at deploying WordPress.com, but I noticed that if you had a new idea and you needed to bring it to life, and it kind of makes sense, like they evolved to host these really important .coms. And so they hadn't thought about that zero-to-one experience for any new idea that a developer would have. And so when I realized that, it dawned on me, I couldn't wait to leave and say, this could be a huge business opportunity because the cloud was starting to take off and the cloud was known also for its reliability, robustness, its scale, et cetera. But it felt like the opposite of using a new computer that is like just point and click. It was this really difficult process of configuring and figuring out instances and And so sort of in parallel, it was this convergence of ideas, make developers really productive and happy, give them the best possible tools. I built Next.js, but also figure out that cloud part that a lot of companies would put off until too long. Like you get to the point where, you know, I meet companies that are operating at massive scale that are afraid of deploying. Companies that deploy once a quarter, companies that do code freezes for weeks, months around Black Friday and things like that. So This was not obvious to me. I kind of discovered that when I started Vercel, but I thought it was all going to be about the instant feeling for developers, which is a big part. But then as I started talking to enterprises, it was this idea that as they become bigger, they become stagnant, they become ossified in their old infrastructure choices. And you could imagine now with AI, that problem is 10x in terms of importance. Like we have models coming out every day. There's just like 2 new open source models last week. Imagine a company that is ossified in its deployment and infrastructure choices.

When you started Vercel, because you had this background, did you have most of the insights that turned out to be true about what Vercel was, or was it much more of, you know, the problem area, you know, some of these customers, but you figured it out as you went?

I had a lot of things that were based on first principles thinking. So literally measuring the speed of light and how fast can you shuttle files on a laptop to a container in the cloud that starts a build process. And how fast is that build process going to be? It's funny, I was listening to DHH and Lex Friedman, and he was saying the gold standard of DX was set in the '90s by PHP. PHP is a technology that only required of you to modify files in a folder. Imagine Dropbox, if Dropbox was a live deployment. You go into the magic folder, you edit a file, and it's live on the internet. And so that was from first principles. I was chasing that dragon. There were a bunch of things that I didn't know. I think Vercel ended up focusing a lot on the frontend side in the beginning of our journeys. I realized that the biggest gap was, or the biggest opportunity I would say, was backend code was really important, but the differentiating experiences of the internet were more frontend heavy. So when you look at something like ChatGPT today, A lot of what it's doing is happening right in front of your eyes in real time. It's thinking, it's streaming, it— you start working on a document and it splits the application in two and there's a live document here. You cannot do those things without powerful frameworks that are focused on the frontend. And that was the alpha in the market. Like if I had come to market with like, again, another runtime like PHP, people would be like, oh, I have to learn a new language and what's in it for me? And so anytime you bring a new tool, it's like a big ask of the world. You have to learn a new thing. You might have to migrate. This is like, it's the M word. It's like you tell an enterprise you have to migrate. So what's in it for me? And what we realized, I actually posted some stats today that were bewildering about a really large retail chain that moved on to Vercel recently. I realized that by giving people the great developer experience,, but without business outcomes, it would also gonna be an incomplete equation. So when you move to Vercel, you now realize business outcomes. Your stuff gets faster, your stuff gets more dynamic, you're able to iterate faster and like build AI products. So that didn't come like overnight, obviously, and it's very hard to bootstrap the 0 to 1, but the idea that it can give you receipts, I think is pretty novel in the cloud itself. The cloud, in fact, in its initial conception has a really bad financial equation because companies have data centers. They've spent tons of money in buildings, in racks, in servers, in storage appliances, in people. And then you ask them, okay, begin the journey to migrate to the cloud. Yeah. Okay, cool. So now I'm going to pay for all of that and concurrently I'm going to pay for the cloud bill. Cool. And what do I get at the end of it? What am I going to see? Right. And so I think what's benefited me is I've always been a really visual design person. So if there's nothing I can see or feel at the end of that rainbow, then this software engineering project is just, you know, kind of pointless to me. And so we focus a lot on actually working backwards from your resulting URL that you deploy in Vercel. Even if you didn't do big A/B test studies and whatever, it's just going to feel better. And then we're going to teach you how to work backwards into the technology so that you start from that strong starting point of user experience. I think it's totally novel in the cloud. Like the cloud has always been about implementation details and primitives and Vercel made it about products and experience.

I think that's a good transition moment because what you talk about here is, you know, starting from the end and then working backwards. And I think a lot of CodeGen is in some ways, you know, there, there's maybe some analogies there. That was a lot about the beginnings. Now sort of fast forward all the way to today. Where I think CodeGen, you're kind of at the center of both the thinking and sort of where the technology will be. Two things are true. One is there are a lot of people that sort of have shared information about how, you know, you go out and you test what developers are actually doing with the products today, and it's not working in such bewildering ways quite yet. You know, like there was a study recently sharing that like people thought they were 20% faster. Turns out they were actually 20% slower, which seems like you know, very surprising in some ways, but you know, we talked about this and it actually wasn't that surprising to you. But on the other hand, there's, you know, probably never been a market, at least that I've seen, that has more excitement, more market pull. Companies are scaling revenue probably faster than anything in this space. And so, um, how do you see the state of affairs like today? Like, you know, like what's going on out at the offices in here like this exact hour?

This is a super hot topic because at Vercel we monitor the idea of landing something. Very closely. Like to us, it's not just about writing the code, is that you land it. And in fact, the higher the bar for what you consider to be landing, the better for you.

Landed means fully deployed and usable?

Even better. Like landing, it means that you've seen some business outcome, like user adoption.

All the way through the funnel.

Right? And again, this is not always easy to do because you also have to balance it with like a healthy dose of like shots on goal, but you have to be disciplined about what it means to land your software. And again, this is just from that principle of always working backwards from like there's some pixels on someone's screen that also will result in some kind of business outcome, new conversions, signups, new information, vibes, whatever you want to call it. But I do encourage people to think about the idea of like landing. And I do think one of the emergent things that we're seeing is there's almost like a bottleneck in landing today. It's almost like code generation has been solved. Everyone's sharing their stats about 50% of my company the code is coming from AI is 80%, 90%. So I do think we have to be rigorous about, well, do you just have a bunch of PRs that nobody is reviewing?

Do you feel that way? Do you feel that code generation is, roughly speaking, solved at this point?

So let me give you a framework of how I think about it. I think on one extreme of the equation, we have vibe coding and tools that cater to a very broad set of users. That's where v0, our agent, lives. And you see products in this category that are saying, you don't need to know how to code. Everybody can cook is kind of our motto, right? So you go to v0, you put in a text prompt, and we create an application. On the other end of the spectrum, you have an engineer that's been working on a 20-year-old codebase, and they want to quote unquote write code faster. They have a solid mental model of how the code works. Obviously, there's gaps in their knowledge. They might be more new to the team, they might be more senior, they might forget certain things. So there's a lot of gaps that AI— I call this AI-assisted engineering or agentic engineering. There's an engineer that's sitting on that chair that is being augmented by AI. And so when you look at it from this perspective, I think, you hear this statistic of more and more of this code bases that people are working on, the code review, the patch is increasingly coming from AI rather than the human. And so the reason I think code is— I can see the line of sight of when code gets fully solved is that we will continue to make the models better. They'll get statistically more aware of all of the patterns in the world. This is almost like a self-fulfilling prophecy, right? The models are learning from all the code that we're writing, from the code that gets rejected, from the code that gets accepted. And so more and more of this code, I think, will straight up come from the agent. The problem that I'm seeing today is that for the most mature engineer projects, it's unclear that the human can fully trust the outcome of the AI, meaning that yes, we're generating a lot of PRs, we're generating a lot of code. The bottleneck has shifted to reviewing that code. Mm-hmm. I've been sharing an anecdote that I heard a couple weeks ago. From a company in the ecosystem where the agent generated a PR. It did exactly what the engineer wanted to do. So this is the mode of thinking where the AI is just reading your mind, but you're giving it a little bit of direction. And so you're like, as an engineer, like, oh, thank God. But the AI deleted a crucial line of code. This happens a lot with models, by the way. You take Sonnet and you realize that it's good at instruction following, but it's not perfect. It kind of goes off the rails just a little bit. In a very heavy-duty, mission-critical production codebase, one line of code can literally mean the distinction between like the whole internet goes offline or everything works perfectly. One line of code could be a cloud infrastructure file that deletes a critical resource or a line of authentication code. You've been seeing some vibes shift a little bit on the is the code that gets generated by AI secure on X and so on. And so when that PR went to review, I think this is where I think a lot of the focus will shift. And we're certainly making a lot of investments here. We need to have agents give you a higher baseline of confidence that the code change is actually good, that it's safe, that it follows best practices,. And so we have agents producing the code. I think now we're gonna have agents reviewing the code. And again, the, the, the human being needs to be there still in the loop.

Does this problem have anything to do with who wrote the code in the first place, or is this merely just a problem that exists for a large codebase? In other words, like a, a, a 30-year-old company that's now trying to use AI that is, you know, making mistakes as it's trying to navigate the codebase. Is that a different thing than a large new codebase today that was generated by AI code, or is it not relevant?

I think so. I think so. One of the things that we do with v0, and again, because we cater to everybody and not just developers, is we're more opinionated. So I think the way that this world is gonna split is that when you're building a vibe coding platform, you're more vertically integrated. Because if you're more vertically integrated, and I'll make this really concrete, v0 produces Next.js code. Vercel builds Next.js. We have integrations into a particular set of database partners. So we can continue to work very, very, very closely on monitoring these known quantities. And actually, I think that if you take that more constrained approach, you could actually exceed what a human being is capable of because I can take all of the best practices of security in the world and I can embed them into the model. And the model can now say, okay, like, I will refuse to do something that's really dumb from a security perspective. Now consider this world where you need the model to be almost like beyond superintelligence. You need it to know the entire history of your business, all of the semantics of the code. It needs to know about runtime interactions of that code, because keep in mind that just working on the codebase is a very incomplete model to go off of. The human is thinking about all of the known data that happens at runtime. Obviously, you kind of, it's kind of sloppy realistically 'cause like you kind of know about it in your brain, you know, oh, our baseline throughput is 1,000 requests per second. So I have to be really careful in the algorithm that I'm writing in this hot path. How does the LM know about that? The LM would actually have to ingest all of the runtime data. Think of it as like piping all the logs and metrics, et cetera, into its brain. And then it also has to know about every possible programming language, cloud platform, deployment tool, et cetera, et cetera. And so you also asked something that I thought was really interesting. You said, is this different for companies that are adopting AI today? And they've been writing code for decades. I do think there's also a human habit delta. I think there's people right now that are being born into this world. Yeah. And so they will learn the ins and outs of how to be careful and wield this tool correctly or not. But I do think that we're all learning as we go, you know, like the jury's still out in like, what is exactly, you know, AI engineering 101? Like there's almost like you still need a curriculum, which kind of defeats the point of the ideal of like, oh, you don't need humans.

Totally. It just seems like if there's like a whole new way to build a house, And you have new people going in and they get to start from scratch and they also know the new way. That's right. Versus an old, you know, house that you need to go completely retrofit and the people working on it are learning the new tactics.

That's right.

Seems like it could be very different.

Totally. And I think what we're seeing from the Next.js side, I actually literally tweeted earlier this morning, like, there should be a convention for agents to advertise to tools that they're agents and not humans wielding the tool. So the tool itself can be polymorphic. It can work, it can adapt itself a little bit differently when it's being wielded by the agent. So if you're building a house and you're a human, I give you the hammer that is adapted for your motor skills and the shape of your hand. But if you're an agent, you're like, awesome. And you're like, from a motor skills perspective, you can hold 20 hammers at the same time. So I will expose a different interface. And this really uncovers an enormous potential that exists today, which is The idea that you can create a whole new set of languages, runtimes, and frameworks that are custom tailored for agentic engineering. And so that goes back to our vision with v0 is that it actually could be the case where almost like a self-driving car, the self-driving car in its version today has some limitations. Waymo cannot go on the freeway still. Still today, right? And it's not in New York, whereas Uber just like took over the whole planet immediately. But Waymo has a lot fewer incidents. It literally does not make a lot of the mistakes that humans make. So I want the tools like v0 to go in that direction where you're almost constraining in a realistic way what it can do. You're not, you know, overselling its potential,, but the things that it does do, it can even exceed what a human would do otherwise by hand. Yeah. So that's the spectrum of flexibility. If you're using, you know, Claude Code or Cursor, anything is possible. Like you can work on any code base. If you're using things like v0, it's more opinionated. It opens the top of funnel to everybody in the planet. And hopefully we can continue to steer it to a place where the output quality It's almost more of a guarantee. And this is actually also what I would encourage everyone to think about with regards to AI. We need to raise the bar where AI is not just a slop generation machine, right? Like we should be thinking kind of like a Waymo, like, oh, look at that in this dimensions, it's way better than what a human would do.

Is your read right now that the overall advantages coming from Claude Code and Cursor and things like that are more self-perceived than genuine?

There's a lot to unpack there. I was speaking recently to a CIO of a very, very large engineering workforce, maybe one of the largest in the world. You know, there is obviously the cultural resistance to new things. So like, obviously, you know, there's always that caveat and whatnot, especially at this scale. But one of the things that he pointed out that resonated was they spent a lot of time evaluating the perceived versus realized productivity gains of AI tools. And he mentioned that the gap between perceived and realized was extremely significant. Meaning that in order to finally roll out the assistance, which they did, people thought this is the second coming of Jesus. This is the— is like best thing that's happened since sliced bread. When you would ask them from like anecdotal perspective, right? But then when you measure what actually gets landed, yes, not so much.

And I bet it's because there's like a multi-step process and in some of the parts now all of a sudden you're going like turbocharged. In other ways you're going slower than you used to and you just focus on the parts where you started flying.

That's right. But also there is the vibes that are changing, right? And this is where I'll make the anti-case of that. I'll say, look, one of the hardest things about software engineering, and I've always, you know, when people would make the joke of like, ha, developers have it so good, they go to Google and they have their lunches made and whatever. That's like a perception that built up with like the HBO series and like maybe the people that have never engineered could be persuaded to think that way because you sit on a You have your nice laptop, you're drinking your smoothie or whatever, but in reality—

There's also the Day in the Life series, you know, those, uh, those little videos of people sharing like—

Yeah, yeah, exactly. That's exactly what people I think would picture, right? Would say, I show up to the office, I make my tea and— Think about that engineer that I just described that feels the weight of the world on their shoulders because that line got deleted. And literally what happened as a result of that is one of the world's most recognizable brands in the world, .com. Went down. And that pressure, that weight, and it's not just even on the runtime and operational aspects, right? It's also, and this is why Vercel, by the way, we worked so hard to eliminate those failure modes because of my own experience of like what it's like to hold a pager.

Totally.

It's one of the worst feelings in the world.

It's a bit like when, I mean, it's not exactly the same thing, but it's, you know, when somebody does some task and gets paid a huge amount for a small amount of time. And they're like, you didn't pay me for that hour.

You paid me for like the 10 years that I got ready for that hour. This is so true. The operational aspect is just so underrated and overlooked. And, but there's also another one, which is the build time aspect, which is, yeah, you are looking at a laptop, but your brain is like literally in pain, just like the hyper-focused, like, you know, Aspie style of like, I can only do one thing at a time and it's a lot of negative feedback loops, error, error, error, error, error, error. And then maybe you get it right. Maybe you could go hours and hours and hours. I sometimes read Slack messages from our engineers like, holy crap, man, I just spent 3 days struggling with this problem. And it was this little thing over here. Like, it's just, it's an emotional roller coaster. And now you're in this new world of like agent do that thing. And now the agent swallows all that nastiness because there's a very important thing that has also happened, which is the world has now hardcore upgraded from assistants to agents. One of my internal things that I repeat is like, agent is a new model. If you pay attention, you know, v0 launched a platform API, which is basically like a very high-level agent that you can automate. You can, if you were Squarespace and you wanted to reinvent yourself, you could use the v0 platform API to create your own website building agent. If you look at Cloud Code, they launch an SDK and people are integrating in various ways. So what an agent is, is that it's the LLM on steroids. It's the LLM focused on an output because the LLM, when it first appeared, it was focused on request response. Write me a poem, translate this to Spanish, translate it back to English, make it funnier, explain it like in 5. So it was all about, you know, you ask for something, you get it back.

Now it's get me an outcome.

Now it's getting me an outcome. And that is absolutely phenomenal because that's how we went from ChatGPT replaces Stack Overflow for asking programming questions to now it's like, build me an awesome website. There's a really funny meme of a really jacked Chad type which is build me something, don't make mistakes. And so going back to the psychology of a developer, think about what that does to you. You went from struggling, this silent struggle that most normies would not know about, to you're just completely in control. And then so we added a sound to v0 when the generation is completed. People have said it's almost like hitting a jackpot because the sound only happens when the agent went through all those errors, all those intermediate states that a human would have suffered with. And now, okay, once we've rendered successfully, we give you the sound. And so now I do think we have to study the psychological and emotional implications of that and also put them into the pile of productivity, because maybe now you feel better and the productivity outcome is unknown.

This is sort of a goofy question, but— and I'm not I'm not, you know, an engineer, but, um, my, uh, my co-founder was, and a lot of good friends obviously have been. What's been described to me is that struggle, that multi-day struggle results in the most, like, you know, satisfying, happy moment at the end. And that that's like a big part of the love of the game.

Yes. Oh, absolutely. Will we lose some— it's like working out.

Will we lose some of that sort of like hunt that some people find?

Yeah, I think it depends.

Maybe it doesn't matter.

It depends on the individual. So v0 can be used by my kids. V0 can be used by PMs, designers, people that are focused on building products. And so they didn't have that feedback loop of like, I'm writing code, I'm analyzing the error. There's still a sense of gratification when you get something that matches your vision, which is really powerful. Oh, now I got— oh, okay. Now I was able to dump my idea into the right prompt and I steer the agent into the right outcome. And there is a huge amount of satisfaction from that. But there is a meta problem here that I do think about, which is that, have you watched WALL-E? Yeah. And like we're floating, like the future is like fat blobs floating and slurping.

At some point, and we're a long ways away and we can talk about it, but like, you know, if all the hard problems take away, do we become that guy?

Yeah.

Just like sitting back enjoying.

Yeah, I do think about that.

And I think everybody enjoying leisure constantly would be a huge problem. I don't think we're going there personally. I'm curious if you do, but I think that would be a huge problem if that happened.

Some people have brought this up, right? Like, uh, when you're solving all of those programming problems, you are building a stronger mental model of the world. You are discovering patterns, acquiring knowledge, you're testing your discipline. I think the discipline part is underrated. Like, I remember my dad, I always remember this one afternoon, I had just gotten my new developer setup. I was super motivated. My bedroom was next to a hallway, and so people would walk past all the time. And I was so locked in, I probably had spent like 7 hours coding nonstop. And I remember two things. My dad walked in one direction and then he walked back. And to me it was like 10 minutes had passed. And he's like, dude, like, how do you do this? You've been coding for hours and hours. And he didn't even know what the output was. He was just impressed by the fact that someone could be so locked in. And so I think programming taught me that. It taught me how to focus. It taught me to be disciplined. It taught me to receive this negative feedback from the compiler and overcome it. And I do think, you know, we'll need to find what the next version of that is, because I don't think it's going to be programming necessarily. Again, like all of— it was so funny. I still am in the game of like, I try to read every user problem, every error message, everything gets reported on X I read. And yesterday I got an escalation from a startup. It was like, I think Vercel is broken. And they escalated through the VCs because they wanted to bypass any support mechanism, whatever. And I read the message and I see the error right there in the message. So it was a complicated stack trace for nerds. It was a 3-linked error objects. So like the stack trace is like error, error, error, but the error was right there. The explanation was right there for why it happened. There was no problem in our— runtime. And again, my posture here is that every problem that a customer has is my problem. In fact, every problem the internet has I see as my problem. I try to like absorb the whole weight of responsibility, but very concretely here, you know, I told the customer like, it's right there, it's this. And in this case it was another vendor that was down returning a 502 error. And so my lesson from there is that I do think that vibes are shifting.

I don't think people are gonna read, And they won't even read a single error log.

So again, going back to my model of like, I'll solve the problem, and then if there's another problem, I'll redefine the problem to be bigger. We are working on an agent that'll compress any metric anomaly. So the PM working on this was super happy to get my forward because we're working on basically, I call it going from problems to solutions. The old cloud was problem, problem, problem, problem. CloudWatch data, blah, blah, blah. The new world is here's the solution, here's the PR, or here's the insight. So it was a kind of epiphany for me of like, we need to double down on that because this complex error message stack was for machines. I can make the case that we should have never given that to a user.

Should have shown like a natural language error in some way. Correct.

Yeah. And in fact, probably what this customer wanted was, and I mean, this is kind of like, maybe too hardcore for today, but here's the email that you should write the CEO of that other company. Like go and complain to that vendor because it's not us. But by the way, I think this is gonna resonate strongly with a lot of people that build platforms and infrastructure. Fault attribution, correct, precise fault attribution is one of the hardest problems in platforms. One thing that would always frustrate me about the internet was people would give a lot of shit to the Google Chrome team, they would say that it consumes too much memory. And that makes absolutely no sense because it's the web pages that consume the memory. How does that nuance not get lost to an end user? People just think— they started memefying it. Like, if you've ever seen this guy that eats a lot of things, it's Chrome eating all the available memory in your machine. And there's things that Google Chrome can do, like they can evict tabs more aggressively, which has its own set of downsides because it's like, Oh, I shut down this app for you, huh? Because I wanted to save memory. And there is, of course, the amount of engineering that it takes to do really good memory compression I don't even want to get into. But fault attribution. So what they ended up doing, I think they still have it, is that if you hover the tab, they say how much that website is consuming and you should take it up to that website. And there's also this really interesting question of like, I love a web that's extremely powerful. The vision of the web put forth by Apple is that the web should only render, like it should all look like Wikipedia. It's like a glorified ebook reader. I love the fact that the web is this insanely powerful platform, but when you give people more power, you get into this fault attribution problem. And the cloud has it in spades because your visitor only cares about the fact that if you go to like, you know, hotwebsite.com and they're going to buy something, it works end to end. But if you have a lot of vendors in the pipeline, then how do I tell you, the developer, oh, it was this company that let you down. It was not Vercel. And so agents can solve this for us. And by the way, it seems like something small and I'm like nitpicking over like this house of cards, but like just telling people concretely what went wrong, it can make a huge difference.

Hearing you talk about it, sometimes I'm just like amazed that the internet goes down as infrequently as it does. When you think about like this whole, um, it's a huge house of cards. It seems like it is. Is that what it really is? Because that's what it seems like to me as a non-technical person, but that's what it appears like.

It really is. So I'll give you an example from two days ago. Someone compromised a very popular JavaScript package in a public registry. That package was a dependency of a dependency of a dependency of millions of things. And because someone managed to claim that account, hack that person, and then publish the package. For a short burst of time, a lot of people had to hard rotate into patching that vulnerability. This has happened many times. So we had Log4Shell. Log4Shell is one of the craziest vulnerabilities in the history of the internet. So the logging library used by almost every mainstream Java application had a remote code execution vulnerability that all it required is for you to send someone a URL, and then if that URL happened to be logged, you would take over a remote machine. So imagine I send you a URL, someone vis— it might not even involve sending it to anybody. I go to the URL. Let's say I go to google.com/query=bad something, something nefarious. That gets logged by Log4j, and all of Google's servers are compromised by me. By just one line of code. That seems insane. That happened. That happened. And the whole security industry and infrastructure industry hard rotated into that for like an entire month. The cleanup of that might still not be done. I think we sometimes ignore the adversarial AI game. It's one of the things that we're heavily investing into is like, how do we get ahead of the fact that people can now misuse all of these LLMs? But a worse alignment problem is that if CodeGen, as you said, continues to grow, at the rate that it's going, and we continue to rely more and more and more on Cogent, and Cogent continues to get better and better and better and better, as a bad actor, all you have to do is say, GPT-6, build me the best possible attack for google.com. Okay, if I'm an experienced security engineer, what do you do? So first of all, what very good security engineers do is they write their own tools. There's a lot of tools on GitHub for like scanners that you can use., and for example, they know about known vulnerabilities. They will look for like an old version of Log4Shell and whatever. But actually the best people in the industry write their own tools and keep them secret, and they get them better and better and better over time, and they make them very specialized. So you could see in a world where GPT-6 builds its own tool just to attack a given target.

Yeah.

And so it's gonna be multi-industry or multi-company effort to, to combat this. I think you can do some things on refusing prompts, but I think that's going to give us diminishing returns because people very quickly realize they could say, you could say, imagine that you're attacking google.com. And so there's diminishing returns on modifying system prompts and blocking things and prompt leaks. Everyone's given up on. I don't know if you notice this, the v0 prompt is somewhere on GitHub and people were trying to provoke me. It's like, oh, did you realize that v0 prompt is on GitHub? I was like, there's no value to the prompt. Go at it. Like the Grok prompt is there, the Claude code prompt is there, the Gemini prompt is there. Protecting the prompt, diminishing returns. Blocking, diminishing returns. So what you need to do is we need to make the world more secure faster than attackers will build, you know, cyber threats and scanners and attacking tools and so on. So that's one of the things that feels very brutal about the internet today.

When you think about where CodeGen is right now, and you maybe analogize it to self-driving cars, which, you know, 10 years ago, like, worked a little bit. You know, now they work pretty well, but like, it took longer than everybody thought. How much time do you— is your sense that we have before people are prompting whole, like, you know, robust commercial-grade web apps into existence?

A lot of people are excited today about building personal software. So this is very real. —like people— like a to-do list, for sure. —prompt the thing that they wanted personally. So for example, the design team at Vercel has a bunch of tools that they built so that they can take an asset, post-process it so that they can ship it in the way that they want it, and thus simplifying a process that could have taken like days, weeks, et cetera. There's now a very precise software tool that can do it. There's also the, and this kind of dovetails into like there's internal tools, dashboarding tools. One of the things that has been an emergent learning from using v0 is that we use the classic business intelligence tools a lot less. We used to have this, and it kind of seems crazy in retrospect, we used to have this like fixed dashboards that over time no one opens. So of course you have the top line company metrics and that like you have your master dashboard that people you know, you have to come back to. You're working on your board deck presentations and you have an authoritative one dashboard maybe. But then all of these BI tools evolved to have this huge list of things that get unmaintained and people don't look at. Because in reality, what people want is generative UI. Mm-hmm. People want to have a question and to produce the perfect visualization for that question just in time. And that's something that agents have already taken over from any other category of software. Meaning if you were to say, can you replicate, I don't know, name your favorite, like Microsoft Power BI, can you replicate it with Vibe Coding? I would say absolutely freaking not. If I have to create a platform that supports 10 million enterprise-grade customers and whatever, it's very hard to compete one-on-one to the entire platform. But the way that they're going to get defeated potentially is more sneaky. Is I don't— I no longer go to that big bloated platform because I can just in time generate exactly what I needed. And so it's a different conception of software. It's not that I was replicating what existed, is I'm shifting the way that I build altogether.

Yeah, it's not that somebody is going to build a new company competing directly with that. It's that each of those customers will just make their own little thing.

And that's already happening. Yeah. And that's— and again, the The rigor of that is even arguably higher. The quality of the visualizations, the component system that it fetches for rendering the data, like all of that looks better, is faster.

They can change it if they want to. Correct. If they want some new features, they add it same day. Yes.

I had an anecdote, so I presented yesterday, so to give you an idea of how v0 is expanding. So a lot of it is PLG, just like Vercel, that you start using v0, et cetera. But this CTO at a company basically started vibe coding in his own personal account when he was talking to a customer. This is a CTO, not a PM, et cetera. A person that would've never written a UI, a frontend code in their life. And so what happened is their customer was saying, you know, if you don't give me this feature, we're basically gonna churn over time. Classic conversation that you get with your customers. What he did is during the conversation, he vibe coded, okay, let me understand exactly what you want. So he vibe coded a version of that feature. The guy in that same conversation, 30-minute conversation, said, that's exactly what I want. So this is the other emergent use cases. You basically built an end-to-end version of what the customer wanted plugged into your data source. And now it can be, if necessary, it can be merged into the mainline platform. So what the guy told me, and this is what he had me present to v0 to the rest of their entire FieldCity organization, is I want everyone to build this way. I want our data to be exposed into this vibe coding platform so we're not left behind. And the announcement was this feature is going live in Q4. But now this is important, right? He did vibe code it in 30 minutes, but to be merged into the mainline monster is not so easy. You don't just vibe code, you know, the production system and whatever. And this is why I was arguing that to merge things into monsters, monster codebases, things with like millions of customers, et cetera, you need slightly different tools, slightly different techniques. And I think that's why the world will continue to partition further into like this highly specialized tools for each kind of problem. Yep. Agents, basically.

What do you think a kid who wants to create software in the future should be doing right now? Because in the past it was more clear, like learn to program and, you know, now let's say you're 16 or 13 or 10, and you know, you're starting to want to think about these kinds of products, what would you spend your time on now?

The best advice I've ever gotten on how to improve my software engineering was have a product idea in mind. If I tell you go and master programming, like it's topology in math, it's weird. Like it's where do you focus your efforts on? What language do you learn? It's like everything is a problem and nothing is a problem. And so what I would say is if you start with something like v0, you're starting to build that muscle of what do I really want? What do I want to build? What am I visualizing in my head? And I do know that this is kind of a somewhat controversial topic because there's a whole line of thinking in psychology that some people can visualize and some people can't. Have you heard about this? Like the visualization power of people is like, uh, The different kinds of eyes in Naruto, like the amount that you can picture in your head varies by person. Yep. But I do think that we have to exercise that muscle. It very well might be that what we call taste is actually that ability to refine that idea that you loosely have in your head and refine it and refine it and refine it and visualize that next state. It's like instead of next token prediction, it's like next, refinement prediction that you're doing in your head of like, this is what it needs to look like. That's interesting for people to absorb it positively.

That was something I was going to ask you about was because, you know, taste became a buzzword and, you know, I think it can, it can have meaning, but it can also be kind of meaningless, but it feels like there's something there. And that, that's actually a pretty good, it's like the best attempt at a working definition I've heard that it's your ability to sort of like visualize a future and you can actually see how that would explain some people's ability to see a trend early or to see a talented person when they meet one or those kinds of things.

I also think it's people focusing on that end of the rainbow that we talked about. Like, how much are you getting distracted by the side quest that went into building the thing versus, okay, what is it going to do when it lands? What does it look like? And I think this is why we created Vercel is that we thought the world has to change its focus a little bit. The cloud has been all about that low level, it's like material science. And yeah, materials are extremely important. I think Vercel can do a good job at curating some of those for you so that you can actually put your focus on the sky. What is that next thing that you're going to give the world? And this, by the way, applies to everything. It applies to how you talk about your thing, the words on the screen, the way— what do you attach to your tweet when you get that thing out? What the blog post that announces your product looks like. And I do want to build a world where people are looking at that. And it was really interesting, by the way. So a couple weekends ago, I had my— you said 13-year-old. I said to my 6-year-old, he walked into my office and he said, what do you do? Like, what are you doing? And I said, because at that very time I was vibecoding, I was like, well, I'm creating software, I'm creating something. And then we got into this conversation, like, could he create a game? And I said, well, what game idea do you have? And he said a soccer game. And so we built a soccer game. I'm gonna share it on X because it's dope. But the hardest thing for, in that process of back and forth with him was he had the idea of the soccer game that was legitimately his own idea. It was really hard to get him to ask me what the next thing, what the refinement of his idea would be. And that didn't come naturally. So In the end, what we ended up building is a game where you dribble the ball and we count and we make it harder as it goes. And it looks really pretty. The pretty part probably came from me. It's my own personal opinion. My lesson there was that, again, you have to build this muscle of understanding there's an audience out there, there's something that technology is capable of, and there's an interest. He was interested in focusing on soccer. And how do you bring all those things together into prompts, into asking the machine what do you want to see in the world? I almost think this is the new— because we quote unquote provocatively have solved code generation, because keep in mind you can go in v0 and you can also ask it to build this data visualization that prevented the churn of a customer for a security company. And so because the platforms can now do all of these things, it boils down to ideas It boils down to capital allocation. Tokens have a cost. So if anything is possible, you know, may the best idea win and it'll get enough tokens behind it that we can keep developing it, right? And so I think that's gonna be the skill of the future. I hope my kids learn it. I hope my kids are able to like do that idea translation really quickly and that mapping.

Does someone still need to understand the lower levels of the code underneath it all?

You know, we didn't get into that at all and we could build a pretty banger game and we could keep going, going, going. And I don't think we would ever look again as the world becomes more agentic.

But like 50 years from now, you know, when all the people who know the deeper levels of code today are no longer in the game, could we have a world where you have a billion people who only know how to vibe code, but nobody knows the stuff underneath?

Well, I mean, we're kind of in that world already, right? Like a lot of people have no clue. How their garbage collector of their programming language works. I mean, a lot of people are pouring effort into interpretability of LLMs because we do not know how they arrive to the answers they do. Right. We will need frameworks. We will need guardrails. We would need ways to interpret and define the semantics. I do think it always boils down not to how the instructions in the processor get written out, and executed, it boils down to that relationship between investment, you know, how much are you willing to pay for a certain level of performance, for example, and outcomes, like what do you want out of this system?

It's just a crazy thought that there's just layers of abstraction and in the future people just won't need to know the, like, ancient layers, but they'll, they'll be driving it all, but Nobody needs to know it.

The world might just be driven by economics, energy allocation. There might be a world in which we've perfectly solved this transmutation of energy source to intelligence, because if this is scaling loss of test time, compute, et cetera, continue, then we pour more effort into the task and we get the outcome that we wanted. Now, the question will be, not everyone will have the same token budget as everybody else. So it really boils down to like who crafts the most compelling vision of the future and the storytelling. By the way, this is a human— this is where Her might have just nailed it. The storytelling is how you persuade the world that your idea is the one that is worth allocating tokens into. Maybe there's a world where like your seed round just gets dumped 100% into tokens and all you're doing is going around the world like, I have this vision of the future. Here's a quick vibe code that I did in v0. But if you really want to to see it at like planet scale. We have to pour a lot more tokens into this thing. And that's how you get to like the one-person billion company and so on.

At the beginning of the conversation, you used a term I really liked, like chasing a dragon. And I think there's like a, there's an interesting debate among founders where some founders believe that you should just have a super clear vision and will it into existence, like creating a movie. There are some founders who are much more iterative and talk to customers and see what it is, and there's pivots and you get there that way. What I observe about you is that you talk a ton to your users. You seem extremely flexible of mind, but then you also use this term chasing a dragon. And so it seems like you've got both of those sometimes.

It's one of the hardest things to do.

Is that how you think about even now as you think about a new product or evolving? Is it this combination of chase a dragon and then listen to get your way there? 100%.

This is the magic of how you build a PLG and enterprise business. The enterprise way of building product has to be extreme. That conversation that I just talked about is like, what do you need? What is your concrete pain points with this technology and like, let's build it together and let's evolve it. But it has to be mixed with the vision of the future because otherwise you're not going to see the next platform shift. You're not going to see what the next generation of coders and builders are excited about. So you have to have that first principle thinking that I sometimes call it internally work backwards from science fiction with a healthy dose of like, well, we're also supporting some of the largest dot-coms on the planet and we have to build things that are very concretely their problems today in the current world today, right? That is the magic of— that's something that can bring success to a lot of people. Now, I'll make the counter case of that. Agents can do their own research. So as v0 becomes more agentic, one of the things that it can do is search the web. So it used to be that humans would say, okay, my boss is asking me to create this really cool animation that does this and that. And so you start sometimes you know a lot of things and that's why we hire experts, but you also go off and do research. And so now think about how agents are just LLM calls in a loop. So an agent is just something that's calling ChatGPT repeatedly in quote unquote. Now think about deep research. Definitely to me, one of the killer apps of AI in recent times, the AI can go off and search 300 different websites and read them for you and then digest them. That'll be wielded by the programming agent. And so in that world, the programming agent can do the user research. And so if you said, okay, I want to build a soccer game, okay, go off and do research about what are users saying in your community forums? What are people saying on X? I can have the PM agent essentially as part of this multi-agent architecture. And by the way, a lot of what I do is talk to users. So I talk to users and say, okay, like search for my keyword and, you know, okay, there's these 10 complaints, let's stack rank them. There's also research with like Salesforce is saying about like pipeline and opportunities, et cetera. So I do think the world will become more and more agentic, but I don't believe that there's going to be one agent to rule them all. This is the promise of what I call the next internet, the transition from HTTP to MCP. If MCP allows us to deploy agents that are highly specialized, then the world becomes a collaboration between agents as opposed to one agent to rule them all. And that is true to my taste for what the web and the internet should be like. I always loved the web because I rejected the idea there's going to be a platform keeper that decides what gets on their platform and what gets off their platform. And controls everything. This is the example of like, people literally were just regurgitating what happened when Elon bought X and then Apple wanted to take X off the App Store. And so we cannot allow that. We cannot allow a world where like people can arbitrarily take you off the internet, right? And so I want a world where MCP wins because that will mean that there's all kinds of agent choices. There's all the companies can rethink themselves as Instead of just having a frontend of pixels, I have a frontend of tools and a frontend to my data. And MCP is emerging as the new frontend instead of for humans, for agents.

Particularly when I was, you know, running Lattice, but just in general now, one of the things I most look up to in companies is extreme coherence when the product and the customers and the culture and the employees all sort of Makes sense together. Yeah. Like Apple is always like an example to me where I'm just like, it's so different. You know, it's secretive. It's got all these different things about it, but it all makes sense together. It seems to me like you with an open source background, with the sort of way you're thinking about the future, being so developer focused, it seems like there would be a culture and a type of person and, you know, the way you run the company that must match that. Yeah. Does that just emerge naturally or is there a way, like, how do you think about making that all?

It emerges naturally, but you have to pay attention because sometimes you just take it for granted. I encourage everyone who runs a business to think about this. Like, I call it reverse engineering your success. Do you know that you can be successful and not fully understand why? I think people don't talk about this enough. All the time. Like—

I think many of the most successful people are like that.

And like, oh, it's because they did A, B, and C.

You might have succeeded in spite of A, B, and C in some cases.

Yeah, and so reverse engineering your success is actually something I devote a lot. It's like, there is a quote about like, It's harder to predict the past than predict the future. Predicting the past means to actually understand— actually came from one of the creators of quantum physics. It's really good. Reversing the past means that you understood the sequence of tokens or like reasoning traces that led you to where you are and you leave out all the noise. And I think for us that actually the culture that we built of openness and transparency— I tweeted the other day that I love working behind the scenes to help our engineers present their own work, meaning I help them. Hey, like I learned these things from X, or like if you say this, maybe you're just being too verbose and like it would look better if you give this example and whatever so that I coach them on how I think about presenting my work. But I want them to own that storytelling. I want them to learn the, oh, this is how I talk to customers. This is how I engage with the community. This is how I receive feedback. As opposed to the traditional thing, which is like you're sheltered and shielded by layers and layers and layers of obfuscation. And you don't get as an engineer the direct signal. But now to your question is, I mean, this sounds amazing, at least to me. Was that something natural? It was emergent, but now it had to get reified into our culture. We need to say that's who we are and that's what we believe in and that's behind our success. But it didn't, that understanding of, oh, there's all these little things that made you successful because some of them are second order effects. Like you said, what I always loved was open source. Well, it turns out that if you love open source, it leads to this open contribution culture. It leads to a higher level of transparency. I got feedback from one of our colleagues that comparing, there's this idea of compare Slack workspaces. How was the Slack that you joined different from the Slack of the company you just left? It's a very important exercise for people to do. And he was saying, well, the Slack at Vercel is just pure information. There's no direct message, thread, group, what— it's just so much information. People learned how to digest what they were working on in a way almost feels like an internal Twitter. And of course companies need a separate, healthy separation between internal and external. And at Vercel, because we're responsible for so many critical workloads, privacy and compliance are extremely top of mind for us. But there is a vibe of like, it's like an internal Twitter in the sense of like, insight, insight, summary of something that I just learned by working with a customer, screenshot. And that is an emergent culture that then you have to say, that's working for us. That's who we are. Let's keep doing it so that when people join, it's not by accident. That's how we work.

One of the things that strikes me about you is not only are you running a big company, you're also deep in the technology, which is amazing to be able to both run the org and know all the technology. I thought I have a lot of kids and they take a lot of time. You really, you really have a lot of kids. I know you're a runner. I'm not like trying to ask like, what's your secret? But like, how do you make it all work? Like, how do you— is there anything that you've picked up over the years through parenthood, through building a company as time pressure has gotten heavy on you, is there anything that you've picked up or adapted that you think is broadly useful?

Yeah, I think there's actually a tweet by your brother Sam that I always loved, which was, you know, it doesn't matter what ladder of success you've achieved, working out is still hard for everybody day in and day out. I do not wanna go to the gym, confronting that thing that I do not want to do every single day. This is why I don't care about what type of exercise, what's hot, CrossFit. I think that's something that's been different about me is like I've never followed any fitness trend. I'm so anti-fitness trend. I have done things that were falling off the popularity cliff. I boxed for many years. I would go every, every day to the gym from 6 AM to 7:30 AM to a boxing gym in the Dogpatch that like the owner would tell me like, I don't know why, like classic Western boxing is not as exciting anymore to people. Because I think it's because MMA has been growing and whatever. I love classical boxing and same with— I love calisthenics. And so I do things that are like off pattern, but it doesn't matter what you do because what you're doing is that confrontation of the thing that you didn't want to do that hardens your mind, stabilizes your mental health, teaches discipline to my kids. My kids love walking into my office. I'm in the Peloton and they're like, holy shit, why are you doing this? They ask me why a lot. Like, why are you doing this? And then I explain that they're not gonna get it for a long time, but they need to learn. And I think everyone needs to learn that like there's something about our own individual personal development that has to do with that confrontation of things that are hard. And that's what I was giving the caveat to, like make sure that whatever relationship you have with that coding agent is one that challenges you and you're learning and you're better for it at the end, right? It's expanding, it gives you more capabilities. You're taking something away from it. You're just becoming the WALL-E guy. So my recommendation is like avoid the WALL-E future. Like you're just like a blob floating in the sky and work hard. And so there's no more secret to that.

The last question I wanted to ask you, back to the taste question and sort of, you know, you described it as this ability to sort of visualize something that hasn't come true yet and how it might iterate and sort of see the path there. Is that something that can be learned or improved upon? And if so, how do you do it?

You know, it links to this idea. I also do a lot of fitness for the meditation benefits. Like, I remember someone, I don't know who told me this, but they're like, look, there's passive meditation, there's active meditation. Passive is you're at the bottom of the tree and you're like, you say your dharmas and whatever, and you go into the nothingness. I see my, you know, really hard workout this morning for 30 minutes of high-intensity interval training as a meditation of sorts. Meditation confers a particular attribute, which is presence. Presence gives you clarity of mind and allows you to capture nuances, reactions, little things about the world. You're just more tuned in. And I do think that one way that you can improve your taste and your product building skills is being more attuned to the actual reaction from people. Because you might say like, here's my thing, and people say, awesome. It's like, is that an awesome of like true joy? Can you be present in a way that reads the world more precisely?

Very hard to have high taste if your brain's off thinking about the next email or the next meeting you gotta go to or whatever.

All of those things, focus, presence, self-honesty. And I think this also comes from that place of confidence and like that med— almost like meditative state of like, can you actually confront the negative reaction or are you running away from it? Because the best thing you could do for your product is to read all of the negative feedback about it. Can you do it? It's almost like running on a treadmill and be like, oh shit. Like my boxing trainer would do this cruel thing. So this is in the Dogpatch for those that don't know, not too far away, maybe like a couple miles away from the, or a mile away from the Giants Stadium, the baseball stadium. And we would work out for an hour and 29 minutes. And then on the last minute he would say, lap to the Giants Stadium and back. And I was like, you just pushed us to, world of pain that is unbelievable. And then you threw another crazy thing on top. And so that was the process of like, there's just always more than you can do. And I do think that there is a level of tolerance of pain that we can always work on and get to greater and greater heights. And you can link this to product building that you have to face and seek that negative feedback from the world. And by the way, a lot of people want to help you. This is actually the nice thing about it. Like, So many people want to help you and want you to build the right thing. And you call this out when, when we talk about enterprises, people really want— they're invested in your success. Yeah. Just don't get in the way of your own success.

Yeah, it's a great place to end it. Guillermo, thanks a ton for your time today. This was great. Thank you.